[Ocfs2-users] Linux kernel crash due to ocfs2

[Sunil Mushran]

Can you provide me with the o2image. It includes the entire fs metadata.
The size of the image file depends on the number of files/dirs.

# o2image /dev/sdX  /path/to/image/file

So the error is clear. We have underestimated the amount of credits
(num of blocks that need to be dirtied in that transaction). This is the most
common write path in the fs and thus hit heavily. So I am surprised by this.

One way to fix it is by reproducing it inhouse. And having the image will allow
us to mount the fs and reproduce the issue. Do you know the size of the write?

On 09/02/2011 07:23 AM, Betzos Giorgos wrote:
> Hello,
>
> we have a pair of IBM P570 servers running RHEL5.2
> kernel 2.6.18-92.el5.ppc64
> We have Oracle RAC on ocfs2 storage
> ocfs2 is 1.4.7-1 for the above kernel (downloaded from oracle oss site)
>
> Recently both servers have been crashing with the following error:
>
> Assertion failure in journal_dirty_metadata() at
> fs/jbd/transaction.c:1130: "handle->h_buffer_credits>  0"
> kernel BUG in journal_dirty_metadata at fs/jbd/transaction.c:1130!
>
> We get some kind of kernel debug prompt.
>
> the stack is as follows:
>
> .ocfs2_journal_dirty+0x78/0x13c [ocfs2]
> .ocfs2_search_chain+0x131c/0x165c [ocfs2]
> .ocfs2_claim_suballoc_bits+0xadc/0xd94 [ocfs2]
> .__ocfs2_claim_clusters+0x1b0/0x348 [ocfs2]
> .ocf2_do_extend_allocation+0x1f8/0x5b4 [ocfs2]
> .ocfs2_write_cluster_by_desc+0x128/0x850 [ocfs2]
> .ocfs2_write_begin_nolock+0xdc0/0xfbc [ocfs2]
> .ocfs2_write_begin+0x124/0x224 [ocfs2]
> .ocfs2_file_aio_write+0x6a4/0xb40 [ocfs2]
> .aio_pwrite+0x50/0xb4
> .aio_run_iocb+0x140/0x214
> .io_submit_one+0x2fc/0x3a8
> .sys_io_submit+0xd0/0x17c
> syscall_exit+0x0/0x40
>
> In the last crash case, the file system was full.
>
> Any clues?
>
> There seems to be a ocfs2 kernel patch some time ago for the 2.6.20.2
> kernel that fixed some journal credits updates.
>
> Is this another bug?
>
> Any help will be greatly appreciated, because this is a production
> system.
>
> Thanks,
>
> George